Baberu OCR Experimental Full-WebGPU ONNX Bundles

This repository contains files derived from Baberu OCR.

Original work: Baberu OCR
Source: https://huggingface.co/genshiai-daichi/baberu-ocr
Original revision: d9cc13153e9a1cd8fdfa3b7b1cc329da2020aeae
Copyright 2026 Baberu OCR authors

Modifications by ameraino11:
- converted constant decoder MatMul and embedding Gather weights to symmetric
  per-output-channel INT8 QDQ;
- replaced the native dynamic-INT8 execution pattern with WebGPU-compatible
  DequantizeLinear and FP32 MatMul/Gather operations;
- exported separate prefill/step graphs and complete shared unified graphs;
- added an INT32 token-ID interface without pruning layers, widths, vocabulary,
  attention heads, KV heads, or decoder weights;
- added browser GPU-resident vision/KV execution, explicit session/device
  cleanup, validation, and benchmark source.

Modified model files:
- variants/webgpu-121/decoder_prefill_qdq_int8.onnx
- variants/webgpu-121/decoder_step_qdq_int8.onnx
- variants/webgpu-121/decoder_unified_qdq_int8.onnx
- variants/webgpu-121/decoder_unified_gather_qdq_int8.onnx
- variants/webgpu-242/decoder_prefill_qdq_int8.onnx
- variants/webgpu-242/decoder_step_qdq_int8.onnx
- variants/webgpu-242/decoder_unified_qdq_int8.onnx
- variants/webgpu-242/decoder_unified_gather_qdq_int8.onnx

The vision_int4.onnx and vision_fp16.onnx files are unmodified upstream
copies. This conversion is independent and is not an official Baberu OCR
release. The repository is maintained as a personal hosting/download mirror
for experimental use and is not represented as production ready.